Innovation Heroes: Why physical AI can’t ship fast and iterate

The AI industry has a favorite mantra: ship fast and iterate. But what happens when your AI operates an 80,000-pound truck on a public highway? Suddenly, iteration looks a lot less appealing.

“In physical AI, mistakes have consequences,” said Tete Xiao, VP of Engineering and AI at Bot Auto. “If you are operating a fleet of robots in a factory, when the robot makes a mistake, the whole line has to be shut down. Same as autonomous trucks.”

And that’s exactly what host Ed McNamara and Tete explored in the latest episode of Innovation Heroes. Tete spent a decade at the frontier of computer vision research — including co-authoring Meta’s Segment Anything model — before joining the autonomous trucking startup.

The stakes are fundamentally higher

When a large language model hallucinates, you get a weird sentence. When a self-driving truck makes an error, the consequences cascade — collision, shutdown, liability. This difference in stakes explains why physical AI remains what Tete calls “broadly underhyped” despite splashy robotics headlines. The technology is advancing rapidly, but deployment requires a fundamentally different approach than the move-fast ethos that dominates software development.

The solution isn’t to slow down — it’s to automate validation. “There’s actually a balance that you can achieve with speed and safety, which is to build better infrastructure and better tool chains,” Tete said. “Sometimes you don’t have to slow down for that. You just need to automate the entire pipeline.”

The biggest challenge isn’t the technology — it’s humans

When asked what makes autonomous trucking so difficult, Tete’s answer was surprising: human drivers. “I would say most of the trouble and the challenges that we’re trying to solve are because of human drivers,” he said. “If everybody follows traffic rules, we’d be fine.”

This insight extends beyond trucking. Any AI system operating in the physical world must account for unpredictable human behavior — workers on factory floors, patients in healthcare settings, shoppers in retail environments. The technology might work perfectly in controlled conditions, but real-world deployment means handling the unexpected.

Why simulation is the cornerstone

You can’t let an autonomous truck experiment on I-10. So Bot Auto runs extensive simulations — scenarios too dangerous to test in real life, from pedestrians on highways to vehicles swerving unexpectedly. Every code change triggers automatic validation against thousands of test cases.

“A couple hours of simulation time is equivalent to years and years of driving in the real world,” Tete explained. This compute-driven approach lets small teams move faster than organizations with thousands of engineers — Bot Auto has just 70 people — because they’ve built the infrastructure to validate at scale.

Indeed, behind every autonomous mile is a massive infrastructure challenge. Bot Auto’s trucks generate enormous amounts of sensor data that must be processed, indexed, and fed back into training pipelines — automatically and at scale. “The magic for pulling this off with such a small team is to have great infrastructure,” Tete said. “Without the good infrastructure, even a big team can’t do much, let alone a small team.”

NEXT STEPS

For IT leaders considering AI deployments with real-world consequences — in manufacturing, healthcare, logistics, or security — the lesson is clear: comprehensive validation, simulation, and infrastructure aren’t obstacles to speed. Listen to the full conversation here to learn what makes speed possible when you can’t afford to get it wrong.

You can also find episodes of the Innovation Heroes podcast on SHI’s Resource Hub, Spotify, and other major podcast platforms, as well as on YouTube in video format.

Video + audio

Audio only